nnvm::compiler Namespace Reference

Typedefs
using	TOpPattern = int
	the operator pattern More...
using	FTVMCompute = std::function< Array< Tensor >(const NodeAttrs &attrs, const Array< Tensor > &inputs, const Array< Tensor > &out_info)>
	Computation description interface. More...
using	FTVMSchedule = std::function< Schedule(const NodeAttrs &attrs, const Array< Tensor > &outs, const std::string &target)>
	Build the computation schedule for op whose root is at current op. More...
using	FTVMAlterOpLayout = std::function< bool(const NodeAttrs &attrs, const Symbol &inputs, const Array< Tensor > &tinfos, Symbol *ret)>
	Modify the op node to alter its input layout. it is invoked in AlterOpLayout pass. More...
using	FTVMVectorizedOp = std::function< nnvm::NodePtr(const nnvm::Node *node)>
	Transform from normal operator to vectorized operator. More...
using	AttrDict = std::unordered_map< std::string, std::string >

Enumerations
enum	OpPatternKind {   kElemWise = 0, kBroadcast = 1, kInjective = 2, kCommReduce = 3,   kOutEWiseFusable = 4, kOpaque = 8 }
	operator pattern used in graph fusion More...

Functions
const PackedFunc &	GetPackedFunc (const std::string &name)
	Get PackedFunction from global registry and report error if it does not exist. More...
tvm::Array< tvm::Expr >	ShapeToArray (TShape shape)
tvm::Array< tvm::Integer >	ShapeToIntArray (TShape shape)

Typedef Documentation

using nnvm::compiler::AttrDict = typedef std::unordered_map<std::string, std::string>

using nnvm::compiler::FTVMAlterOpLayout = typedef std::function< bool(const NodeAttrs& attrs, const Symbol& inputs, const Array<Tensor>& tinfos, Symbol* ret)>

Modify the op node to alter its input layout. it is invoked in AlterOpLayout pass.

Parameters

attrs	The attribute of the original node.
inputs	The input symbols of the original node.
tinfos	The inferred shape and dtype of the inputs.
ret	The replaced operator.

Returns

Whether to replace current operator.

using nnvm::compiler::FTVMCompute = typedef std::function< Array<Tensor>(const NodeAttrs& attrs, const Array<Tensor>& inputs, const Array<Tensor>& out_info)>

Computation description interface.

Parameters

attrs	The attribute of the node.
inputs	The input tensors(placeholders)
out_info	Tensors holding shape/type information about output, & these are always placeholders.

Returns

The output description of the tensor.

using nnvm::compiler::FTVMSchedule = typedef std::function< Schedule(const NodeAttrs& attrs, const Array<Tensor>& outs, const std::string& target)>

Build the computation schedule for op whose root is at current op.

Parameters

attrs	The attribute of the node.
outs	The output tensors.
target	The build target.

Returns

schedule The computation schedule.

using nnvm::compiler::FTVMVectorizedOp = typedef std::function<nnvm::NodePtr (const nnvm::Node* node)>

Transform from normal operator to vectorized operator.

Parameters

node	The source node.

Returns

Transformed vectorized op.

using nnvm::compiler::TOpPattern = typedef int

the operator pattern

Enumeration Type Documentation

enum nnvm::compiler::OpPatternKind

operator pattern used in graph fusion

Enumerator
kElemWise
kBroadcast
kInjective
kCommReduce
kOutEWiseFusable
kOpaque

Function Documentation

const PackedFunc& nnvm::compiler::GetPackedFunc ( const std::string &  name )

inline

Get PackedFunction from global registry and report error if it does not exist.

Parameters

name	The name of the function.

Returns

The created PackedFunc.

tvm::Array<tvm::Expr> nnvm::compiler::ShapeToArray ( TShape  shape )

inline

tvm::Array<tvm::Integer> nnvm::compiler::ShapeToIntArray ( TShape  shape )

inline